Objective Conventional reconstruction methods are insufficient for the reconstruction of seismic data with missing consecutive traces, producing a negative impact on subsequent processing accuracy. Hence, this study proposed CU-Net++, a deep learning network based on the U-Net++ architecture combined with the convolutional block attention module (CBAM).

Methods During the reconstruction of missing data, the independent decoder for each sub-U-Net in the nested U-Net++ architecture enables the utilization of information from different depths. The long and short skip connections can effectively enhance the network's capability to extract multi-scale features from data. The core innovation of CU-Net++ is the introduction of CBAM, which can enhance the capacity to learn about seismic wave details and edge information, into the U-Net++. This helps improve the network's ability to identify and capture complex seismic wave characteristics. Through the reconstruction tests of simulated and measured data, this study presented a comparative analysis of the reconstruction effects for missing seismic data of the CU-Net++, U-Net++, CU-Net, U-Net, and curvelet-domain projection onto convex sets (POCS) methods from the perspective of F-K spectrum, residual profile, single-trace waveform, mean absolute error (MAE), signal-to-noise ratio (SNR), and peak signal-to-noise ratio (PSNR).

Results and Conclusions CU-Net++ delivered the optimum overall performance across various assessment metrics, yielding the lowest reconstruction error. Compared to U-Net++, it reduced the MAE by approximately 51% and improved the SNR and PSNR by 5.87 dB each. Notably, CU-Net++ enables high-precision construction of seismic data with a proportion of consecutively missing traces not exceeding 12%.